Demis Hassabis, Father of Google's Artificial Intelligence

Part 10. Humanity's Future and AI's Responsibility

29 Is AI a Tool or a New Species?

Kim Kyung-ran, Kim Kyung-jin

Innovation and Ethical Tensions Inside a Giant Corporation: The AI Ethics Debate Within Google and Hassabis's Position. In January 2014, on a day when cold winter winds swept through London, Demis Hassabis was making the gamble of a lifetime. What lay on the negotiating table with Larry Page was not just an acquisition price. Hassabis demanded safeguards that would let Google, that vast empire, swallow DeepMind without being able to digest it.

That safeguard was the establishment of an Ethics Board. He was convinced this technology would wield the most powerful force in human history, and he did not want that force swung solely for shareholders' benefit. At the time, he stated flatly: "We are not doing this research just to build a smarter search engine." It was like David putting on Goliath's armor while still holding on to his sling.

But ten years later, in 2024, the situation had grown far more complicated and precarious than Hassabis had anticipated. The generative AI war sparked by ChatGPT would no longer allow Google DeepMind to remain in its ivory tower. At Google's Mountain View headquarters, a "Code Red" was issued, and DeepMind researchers in London could no longer devote themselves solely to writing papers.

They now had to become the "Engine Room" responsible for the survival of the great ship called Google. The ethical tensions that arose in this process shook DeepMind's identity to its core. The original Ethics Board remained shrouded behind a veil, its membership never disclosed to the public, and the friction between Google's internal Ethical AI Team and senior management exploded with the firing of Dr. Timnit Gebru.

Hassabis appeared to stay silent through this turmoil, but internally he was walking a tightrope between his ideals as a researcher and his reality as an executive. Instead of public criticism, he chose to strengthen internal "alignment" research. Inside DeepMind, a self-mocking voice asking "Are we building a monster?" clashed fiercely with the pragmatic argument that "If we don't build it, someone more dangerous will."

The Gemini model's image-generation error in 2024 was the defining incident where these tensions broke the surface. The error, caused by an aggressive attempt to diversify the races of historical figures, was less a technical defect than a "paradoxical failure" in which excessive ethical correction itself produced bias. Hassabis explained that "this was not what we intended," but this was not a problem that could be fixed by tweaking a few lines of code.

It was a symbolic scene showing what happens when DeepMind's ideal of building "universal, neutral intelligence" steps into the mud of Silicon Valley's Culture War. Hassabis's position has shifted in subtle ways. Early on, he insisted on "keeping AI as an independent research body, separate from Google," but he has now integrated DeepMind under Google's company-wide slogan of "Responsible AI."

Some critics say "the idealist Hassabis has knelt before capitalism." But those close to him tell a different story. Hassabis came to a bone-deep realization that without Google's massive infrastructure, reaching AGI (Artificial General Intelligence) would be impossible. At the 2025 Davos Forum, he said this:

"Safety is not a brake that slows you down. It is a steering wheel that lets you go farther." In other words, he began to accept the ethical debates inside Google not as a constraint but as an essential control system to keep AGI from running out of control. He is, in effect, piecing together the hardest puzzle only he can solve, caught between a giant corporation's profit motive and the safety guardrails humanity needs.

The Open vs. Closed, Competition vs. Cooperation Dilemma. In 2016, when AlphaGo defeated Lee Sedol 9-dan, DeepMind published the full paper explaining how it worked in Nature. Developers around the world cheered, and it became the catalyst for a chain of advances including AlphaGo Zero and AlphaZero. Hassabis's philosophy at the time was clear:

"Science advances when it is shared." Carrying on the tradition of the seventeenth-century Royal Society, he believed it was a scientist's duty to put discoveries before the world rather than hide them. AlphaFold, too, was released free to biologists worldwide and helped solve protein-structure problems that had resisted solution for decades. It was a triumph of scientific altruism.

Starting in 2023, the atmosphere turned cold. When OpenAI decided not to disclose the technical details of GPT-4, the age of "Open" faded and the age of "Closed," or rather "Secrecy," arrived. Google DeepMind could not resist this current either.

In a 2024 interview, Hassabis said: "Now, releasing dangerous information can itself be an irresponsible act." Where publishing a paper had once been a contribution to humanity, releasing the blueprint of a powerful model approaching AGI was now seen as no different from posting nuclear-weapon instructions on the internet. This dilemma cut Hassabis to the bone.

He was, by instinct, a scholar who gravitated toward openness and collaboration, but reality was forcing him to raise walls and dig moats. In early 2025, when Meta's Yann LeCun argued that "only open-sourcing AI makes democratic control possible" and released the Llama models, Hassabis fell into a deep silence. He worried that Meta's approach "could create irreversible risks," yet he also had to manage the backlash from researchers who resented Google DeepMind becoming an increasingly closed fortress. Watching many talented people leave, saying "We came here to share scientific discoveries, not to boost Google's stock price," and migrate to the open-source camp was a source of great pain for him. The dilemma of competition versus cooperation was tightening around him as well.

At the AI Safety Summit held at Bletchley Park, England, in 2023, Hassabis stood alongside Sam Altman and Dario Amodei and pledged "international cooperation for AI safety." On the surface, it was a beautiful picture of rival CEOs joining hands over humanity's future. But behind the stage, fierce battles over talent recruitment and chip procurement were raging.

At Davos in 2026, Hassabis sat down with Anthropic CEO Dario Amodei to discuss "self-improvement loops in which models teach each other and advance exponentially." The two agreed they must cooperate on safety, yet neither hid the competitive drive over who would complete that loop first and seize "the last invention." Hassabis has now revised his longtime belief in "open science" and established a new principle: "Staged Release."

He calls this not a scientific retreat but "strategic patience." He believes that when competitors recklessly accelerate, DeepMind, at least, must advance slowly but surely, maintaining scientific rigor and safety. But the market does not wait. Investors clamor, "Why isn't Google shipping faster?" The contradiction between the "open scientist" and the "closed entrepreneur" is the paradox of Goliath that Hassabis must face in the mirror every morning.

Balancing Technological Optimism with Caution: "Technology is neutral, but its use is not." In December 2024, the day he received the Nobel Prize in Chemistry at the Stockholm Concert Hall, the expression on Demis Hassabis's face was closer to solemnity than joy. In his acceptance speech, he captivated the audience by describing the medical revolution AlphaFold would bring. "If we solve intelligence, we can solve disease, the climate crisis, and the energy problem."

The world of "Radical Abundance" he described felt vivid enough to touch. It was the world he had dreamed of at age twelve, staring at a chessboard, where intelligence saves humanity. For Hassabis, AI was neither good nor evil.

It was a neutral energy, like fire or electricity, capable of propelling civilization to the next stage. But the reality he faced after stepping off the stage was not so simple. He knew better than anyone that even if the tool he built was neutral, the humans who held it never were.

In 2023, when he signed the "Statement on AI Risk" warning of extinction-level danger, many people were puzzled. Criticism poured in: "If you're warning that the technology you're building could drive humanity to extinction, why don't you stop developing it?" Hassabis answered: "Even if we stop, this technology will not disappear.

The responsible ones among us must stand at the front and point the way." Hassabis's technological optimism is not blind faith. It is a "conditional optimism" built on rigorous calculation and simulation. He sees the essence of technology as "the expansion of intelligence."

By supplementing the biological limits of the human brain with silicon-based intelligence, we can unlock the secrets of the universe faster, he argues. For example, DeepMind's AI solving plasma control, one of the key challenges in nuclear fusion, was powerful evidence for him. He says: "If we only fear AI, we are voluntarily giving up the chance to save patients suffering from cancer." In his view, halting AI development is itself the unethical act.

But at the same time, he remembers painful history. When Alfred Nobel invented dynamite, he believed it would end war. When Oppenheimer built the atomic bomb, he hoped it would bring peace.

Hassabis desperately hopes he will not become the Oppenheimer of the twenty-first century. That is why, even as he emphasizes technology's inherent neutrality, he takes an extremely cautious stance on the way that technology is deployed into society. He insists that "technology does not exist in a social vacuum" and that technologists, philosophers, sociologists, and policymakers must come together to write AI's user manual. His optimism is not the complacency of "technology will fix everything" but something closer to a willed hope: "If we use it wisely, technology can save us."

Preparing for Bad Actors. In February 2024, at a closed session of the Munich Security Conference, Hassabis was briefing intelligence chiefs from around the world on the worst-case scenarios AI could bring. What he put on the screen was not a robot army out of Terminator. It was a far quieter and more lethal set of threats: biological terrorism and cyber collapse.

Hassabis warned that a tool like AlphaFold, which in the hands of a well-meaning scientist becomes a key to drug development, could in the hands of a terrorist become a blueprint for engineering a deadly virus. He called this an "Asymmetric Threat": a situation where a small number of bad actors, aided by AI, gain the ability to hold the entire world hostage. This is the most concrete and imminent fear that keeps Hassabis awake at night.

To counter these threats, he proposes two strategies. The first is "Technical Containment." Before releasing a model, he puts it through thousands of simulated hacks and misuse scenarios via a Red Team.

The goal is to install strong guardrails from the training stage onward so the model will not reveal instructions for making toxic substances or write cyberattack code. Google's Secure AI Framework, announced in 2025, is the direct embodiment of Hassabis's philosophy on this front.

He has gone so far as to say "An unsafe AI is not AI at all," establishing a baseline that prioritizes safety over performance. The second strategy is "Institutional Oversight." Hassabis argues that a powerful international AI monitoring body, modeled on the International Atomic Energy Agency (IAEA), is necessary.

He believes that "no single company and no single nation can control this technology alone." The "Frontier AI Safety Pledge," in which he played a leading role at the 2023 UK AI Safety Summit, was a first step toward building such institutional mechanisms. He maintains hotlines and shares information on safety issues even with competitors like OpenAI and Anthropic.

Hassabis's greatest fear is that the speed of technological progress will outpace the speed at which defenses can be built. He calls this "the Pacing Problem." Run too fast and you crash; go too slow and you miss technology's benefits. He argues that before bad actors get their hands on AI, we must finish building a "Shield AI" capable of detecting and defending against them. In a 2024 podcast, he confessed: "We are racing against time right now. But the finish line of that race is not technological completion. It is humanity's safe survival."

Ultimately, the future Hassabis envisions is not a world where AI replaces humans but a world where humans, carrying the powerful shield of AI, defend themselves against disease, poverty, and malicious threats. Just as he reads his opponent's moves on a chessboard and prepares his defense, he continues, at this very moment, to play a strategic game against unseen bad actors.

Cooperation Between Humans and AI